Quantifying task-relevant representational similarity using decision variable correlation

TL;DR

The paper introduces decision variable correlation (DVC) as a task-relevant metric to compare neural representations in brains and deep networks, distinguishing task-driven similarity from global representational overlap. By decoding decision variables with linear readouts and correcting for noise, DVC quantifies image-by-image consistency of two observers' decision strategies. Across monkey V4/IT and various networks, brain–brain similarity rivals model–model similarity, while model–brain similarity declines as ImageNet accuracy rises, and adversarial or data-rich training fails to improve brain alignment. The findings imply a divergence between primate ventral pathway representations and those learned by classification-trained models, urging task-centric methods and training paradigms that better capture brain-like, task-relevant representations. The work also shows DVC can complement Cohen's Kappa and RSA by isolating task-relevant decision dimensions and reducing biases inherent in decoders.

Abstract

Previous studies have compared neural activities in the visual cortex to representations in deep neural networks trained on image classification. Interestingly, while some suggest that their representations are highly similar, others argued the opposite. Here, we propose a new approach to characterize the similarity of the decision strategies of two observers (models or brains) using decision variable correlation (DVC). DVC quantifies the image-by-image correlation between the decoded decisions based on the internal neural representations in a classification task. Thus, it can capture task-relevant information rather than general representational alignment. We evaluate DVC using monkey V4/IT recordings and network models trained on image classification tasks. We find that model-model similarity is comparable to monkey-monkey similarity, whereas model-monkey similarity is consistently lower. Strikingly, DVC decreases with increasing network performance on ImageNet-1k. Adversarial training does not improve model-monkey similarity in task-relevant dimensions assessed using DVC, although it markedly increases the model-model similarity. Similarly, pre-training on larger datasets does not improve model-monkey similarity. These results suggest a divergence between the task-relevant representations in monkey V4/IT and those learned by models trained on image classification tasks.

Figures (13)

• Figure 1: The computational framework of decision variable correlation (DVC) for neural representations. (a) Traditional signal detection theory models how a single observer solve a binary classification task. The idea is that the observer use a decision variable together with a criterion (dash line) to make a choice. (b) Decision variable correlation generalizes the signal detection theory to study the trial-by-trial consistency of the decision variables of two observers. The two panels illustrate two cases with the same accuracy in solving the task, but with drastically different correlations in the decision variabless (DVs). (c) We further generalize DVC to compare two neural representations. The basic idea is to use optimal linear classifier to infer the decision variables of individual observers and then quantify the consistency of the decision variables.
• Figure 2: Results on models trained on ImageNet-1k. (a) Heatmap: DVCs inferred for pairs of models. Different colors are used to indicate models from different model families. 15 models are selected to represent this cohort in later analysis. (b) 2D t-SNE embedding of the models using their dissimilarities, measured as $1 - DVC$. (c) There is a strong negative correlation between the classification performance (top-1 accuracy) of a network and its DVC correlation to the V4/IT representation. (d) Networks belonging to the same family exhibit higher DVCs compared to those belonging to different model families (p = 1.33e-56).
• Figure 3: Results on robustly trained deep networks on ImageNet-1k. (a) Networks based on adversarial training has lower DVC with V4/IT compared to the representative models (introduced in Fig. \ref{['fig:fig2']})) without adversarial training. (b) Heatmap showing the inferred DVCs between pairs of models. (c) Robustness networks have high DVCs among themselves, and they have relatively low DVCs with the representative models.
• Figure 4: Results on deep networks trained on richer datasets. (a) Networks we examined that were pre-trained on richer datasets exhibit lower DVC with V4/IT compared to the representative models (trained on ImageNet-1k). (b) Heatmap showing the inferred DVCs between pairs of models. (c) DVCs between lower data-rich models and representative models are generally lower than those within representative models or data-rich models.
• Figure 5: Comparsion to Cohen's Kappa. (a) Heatmaps showing the DVC and Cohen's Kappa for pairs of representative models. (b) There is a strong positive correlation between Cohen's Kappa and DVC on model-model consistency (evaluated on this dataset) (c) There is a decent positive correlation between Cohen's Kappa and DVC on model-monkey consistency. (d) The values of Cohen's Kappa between models (blue) and human subjects (green) are low while Cohen's Kappa between models and between human subjects are high, consistent with the original report. Re-analyzed based on data from geirhosImageNettrainedCNNsAre2018. (e) Scatter plot showing the relationship between Cohen's Kappa and the difference in accuracy of pairs of observers, and the theoretical upper bound. (f) The response histogram to 'edges' distortion based on the model and decision rules used in geirhosAccuracyQuantifyingTrialbytrial2020 and the original study geirhosImageNettrainedCNNsAre2018. Different colors represent different nerual networks. (g) Simulation results show that Cohen's Kappa is sensitive to decision biases, while DVC is invariant to decision biases.